 |
Pharmacist Tien (Pham Minh Huu Tien) launched an AI called ChatDST. Photo: DST . |
On July 14, Pharmacist Tien posted a video introducing many of the company’s products on social media. Among them, the AI chatbot named ChatDST attracted attention with the presenter’s statement that it was “better than ChatGPT.”
This information is controversial because Mr. Tien and the organization behind him work in the cosmetics and entertainment industry and have no expertise in technology. Meanwhile, OpenAI's ChatGPT, which has received hundreds of billions of dollars in investment and many years of development, is still a market-leading product in many aspects.
Statement of Pharmacist Tien
“Almost all of Tien's parameters surpass the artificial intelligence that you are currently using. Tien will reveal to you that this parameter is quite high in the market,” Pham Minh Huu Tien (Pharmacist Tien) declared at the launch.
 |
Parameters of ChatDST chatbot given by Pharmacist Tien. Photo: Pham Minh Huu Tien. |
The example this person used for comparison was the VRAM capacity. Pham Minh Huu Tien claimed that the system they built had twice the memory of ChatGPT (960 GB compared to 480 GB). “Besides, it also has your personality, loaded into the data in the chemical and cosmetic industry, which will be very helpful for those working in this field,” said Mr. Tien.
In addition, the parameters on the presentation screen also show that this platform can generate 2,000 tokens per second, the architecture combines 6 AI models to increase performance. ChatDST is committed to stable operation for 99.95% of the time, processing contexts of up to 1 million tokens.
In particular, the introducer also said that this chatbot has unlimited memory for each session. Pharmacist Tien announced that this solution will be launched in October, users will have to pay to use it, the price is 20-300 USD . However, the launch event video has now disappeared from YouTube.
Abnormal parameters
For now, the project that Pharmacist Tien announced is still “on paper”, without a trial version or detailed reports on source code or operation. The current information only includes basic parameters, along with the statement that ChatDST has a higher configuration than ChatGPT, among the best on the market.
However, this information is also vague, largely illogical and impossible.
Talking to Tri Thuc - ZNews , Mr. NB, Master of Information Technology, AI product developer, said that there was some erroneous information in Pharmacist Tien's statement.
 |
OpenAI's $40 billion AI factory is under construction in Texas, USA. Photo: Data Center Dynamic. |
For example, Mr. Tien said ChatDST was developed on the Llama platform with 403 billion “neurons”. In fact, large language models use parameters to talk about data domains instead of “neurons”. At the same time, Llama was developed by Meta, which released open source code, and only had a 405 billion parameter version, not a 403 billion version.
On the other hand, there is no source showing that OpenAI uses a server with 480 GB of memory as Pham Minh Huu Tien announced. According to the parameters given by Pharmacist Tien, ChatDST runs on a platform of 12 Nvidia A100 80 GB GPUs, with a total of 960 GB of VRAM. The market price of each of these graphics cards is around 700 million VND, equivalent to an investment of 8.4 billion VND for this component.
Meanwhile, a source from The Information said that OpenAI owns a data center containing about 350,000 A100 cards for inference. ChatGPT alone is using 290,000 GPUs, 24,167 times the configuration that Pharmacist Tien announced. In addition, this company also rents Microsoft's Azure servers to reduce the load when the number of users increases.
It is estimated that OpenAI has to pay about 100,000 USD ( 2.6 billion VND ) for each day of operation. Thus, the operating cost for 4 days of this company is more than the amount for the configuration of 12 A100 chips that Pharmacist Tien used at ChatDST.
 |
The Llama 405B takes up at least 810GB of VRAM to house the working model. Photo: HuggingFace. |
According to HuggingFace , Llama 405 billion parameters need at least 810 GB to store. Thus, Mr. Tien's server only has 150 GB free to serve calculations when users chat. This amount of VRAM is enough for 1-3 customers to chat at the same time. Even when reducing the accuracy of the model to save resources, the number of users serving 3-9 accounts at the same time.
ChatDST's processing speed of 2,000 tokens/s is also an "unimaginable" parameter in the industry. Currently, ChatGPT only produces 50-100 tokens/s depending on the model and overloads the system. Claude also only meets 40-80 tokens/s.
Source: https://znews.vn/thong-so-khong-tuong-tu-ai-hon-chatgpt-cua-duoc-si-tien-post1569417.html
Comment (0)